Combining positive unlabeled and active learning for electrical insulator defect detection with incomplete annotations

This paper focuses on detecting insulator defects in high-voltage power systems by introducing deep learning and computer vision technologies to address semi-supervised object detection with limited annotations. To solve the issue of the limited expert annotation, this paper proposed a joint PU-AL (Positive Unlabeled-Active Learning) approach for insulator defect detection. This approach integrates PU learning into the Active Learning framework. In detail, a loss prediction branch is introduced to measure the information content of unlabeled samples. Therefore, the predicted loss is used to select the samples with higher information content. These samples are then manually annotated by experts, thereby improving the detection performance of the network. Experimental results demonstrate the effectiveness of the proposed annotation method compared to random annotation.